Detector-agnostic performance benchmarking for Key4hep simulations
👉 Open the live dashboard  ·  📖 Read the docs

---

k4Bench measures where the time and memory go in DD4hep / Geant4 detector simulations run through ddsim in the Key4hep stack.

Point it at any DD4hep compact geometry and it will tell you how long a simulation takes, how much memory it needs, and — crucially — which subdetector is responsible. It does this without you editing a single XML file or recompiling anything.

What it does

• ⚡ Geometry sweeps — automatically run a baseline, then re-run with each subdetector removed (or only a chosen subset kept) to measure each detector's cost. The original geometry is never touched.
• ⏱️ Per-event & per-detector timing — C++ Geant4 timing plugins record per-event wall time, RSS memory, and per-subdetector stepping time.
• 📊 Analysis & dashboard — load results into pandas, plot them with the bundled helpers, or browse historical trends across Key4hep releases on the live dashboard.
• 🔭 Detector-agnostic — works on any DD4hep compact XML. FCC-ee detectors (ALLEGRO, IDEA) are the worked examples and nightly-CI targets, not a limit.

Quick start

The recommended install is from source, so the C++ timing plugins are built and you get the full set of metrics:

# 1. Clone the repository
git clone https://github.com/key4hep/k4Bench.git
cd k4Bench

# 2. Source setup.sh to source Key4hep, make a CVMFS-aware venv, install deps,
# build the timing plugins, and install pre-commit hooks.
source setup.sh

# 3. Install the k4bench command (editable)
pip install --no-build-isolation -e .

# 4. Benchmark a geometry (single particle-gun run)
k4bench --xml $K4GEO/FCCee/ALLEGRO/compact/ALLEGRO_o1_v03/ALLEGRO_o1_v03.xml \
        --events 100 \
        --ddsim-args="--enableGun --gun.particle e- --gun.distribution uniform"

Want to know each subdetector's cost? Add --sweep:

k4bench --xml ALLEGRO_o1_v03.xml --sweep \
        --ddsim-args="--enableGun --gun.particle e- --gun.distribution uniform"

Results print as a summary table and are written as CSV (plus per-event / per-region JSON) under logs/<geometry>/.

Also on PyPI: pip install k4bench --no-deps (inside Key4hep) gives you run-level metrics, but not the C++ timing plugins — so per-event and per-detector timing are unavailable. Installing from source is recommended.

Analyse and plot the results

The bundled analysis helpers load a run directory into pandas and produce ready-made Plotly figures:

from k4bench.analysis import load_results, plot_run_overview

df = load_results("logs/ALLEGRO_o1_v03")        # one row per run
plot_run_overview("logs/ALLEGRO_o1_v03").show()  # bar charts across runs

Documentation

Full documentation — installation, every CLI option, the sweep modes, the timing plugins, the architecture, and the dashboard — lives at:

📖 https://key4hep.github.io/k4Bench/

I want to…	Start here
Install and run my first benchmark	Getting started
Understand sweep modes & options	User guide
Understand how it works	Architecture

License

Distributed under the terms of the LICENSE in this repository.